fix: normalize Skill tool args to prevent TypeError when models pass objects

[linhongyu510]

Problem

When a model calls the Skill tool with args as an object instead of a string, the tool crashes:

Error executing Skill: replace() argument 2 must be str, not dict

Observed tool call (from a saved session, mr_sessions/session_latest.json):

{"name": "Skill", "input": {"name": "recall", "args": {"args": "--corpus \"$MEMORY_CORPUS\" --k 5"}}}

_skill_tool does args = params.get("args", "") with no type check and hands the value to substitute_arguments, which calls prompt.replace("$ARGUMENTS", args) → TypeError. The raw traceback is surfaced verbatim to the model, which retries the identical call and eventually gives up on the skill and hallucinates the answer (observed: 4 identical retries with gpt-oss-120b via llama-server). Small/open models on custom providers wrap tool input in an object fairly often because every tool appears to them as input: {...}.

Fix

Add _normalize_args() in cheetahclaws/skill/tools.py and call it at the top of _skill_tool:

• str → unchanged.
• dict with a single string value (the observed {"args": "..."} shape) → unwrap the value.
• list of strings (e.g. ["--env", "prod"]) → join with spaces.
• Anything else (multi-key dict, int, None, …) → return a concrete Error: Skill 'args' must be a string… message with a suggested shape, so the model gets something it can act on instead of a traceback.

Tests

4 new tests in tests/test_skills.py (all fail on main, pass here):

• dict with a single string value is unwrapped and the skill runs;
• a list of strings is joined and the skill runs;
• a multi-key dict returns an error containing "must be a string" (no crash);
• a non-string (int) returns an error containing "must be a string" (no crash).

The shared skill_dir fixture now also writes the existing deploy skill (the ARGS_MD fixture constant already defined in the file) so the new tests exercise a prompt containing $ARGUMENTS; test_load_skills count assertion updated 2 → 3 accordingly.

Verification: tests/test_skills.py 30 passed (26 pre-existing + 4 new). Full non-e2e suite: 2782 passed; the 5 failures and 33 web-api import errors are environment-related and reproduce identically on main (macOS rlimit/case-sensitivity, missing SQLAlchemy for the web extra). ruff check on the two touched files reports only the 6 pre-existing findings — no new lint issues.

Honest caveat: the tool's agent run is mocked in tests (_fake_run), so the fix is verified at the unit level; no live model run was performed.